Deep learning approach with optimization algorithm for reducing the Training and Testing time in SAR image Detection and Recognition

Abstract

SAR images have become more popular in the fields of remote sensing and satellite technology. This SAR image can be acquired in various weather conditions, including day or night, cloudy or sunny. SAR images are used for a variety of purposes in image processing, including resource management, agriculture, mineral exploration, and environmental monitoring. The useful information of the SAR image also were affected with speckle noise. Sometimes SAR picture noise is suppressed by a noise deletion by using the filter algorithm on the picture and further analysis prior to display. To do so, the Median, Guided Filter (GF), Lee, Box, Adaptive, or Wiener filter algorithms were utilized, and their PSNR, SNR, and MSE results were compared. GF outperformed all other algorithms in the high PSNR value of 37.8342. Image separation is a necessary step in image processing. Segmentation or separation is used to rationalize and change an image's display into something more relevant and understandable. The character of Hue, Intensity, Saturation (H, I, S) were applied to acquire the information of the pixels of the target image. In this, color information and edge extraction are the basic idea about to achieve the image segmented from its background. Feature extraction can be done in three stages with DNNs: low, middle, and high level feature extraction. In low level the image edge and lines are extracted. Because it's a basic image feature, those all joined to generate a high-level feature. Despite the fact that the training and testing time for SAR image detection and recognition is extremely time taking [1]. For reducing the training and testing time of SAR image, optimization algorithms such as Stochastic Gradient Descent with Momentum (SGDM), RMSProp and Adam optimization methods are used. Their performance shows that the proposed model structure with SGDM optimization algorithm achieved best performance. The detection aims to locate the presence of objects in an image with a bounding box on the region of interest. In SAR image recognition, pre-trained CNN models like ResNet-50, AlexNet, VGG16, and the proposed models was used. The performance of the all three pre-trained models and the proposed models were compared in accuracy and speed. The AlexNet, ResNet-50, VGG16 and proposed models achieved accuracy of 89%, 92%, 86% and 95% respectively. In speed, proposed CNN model with SGDM out performs in training 26’ and 49s and testing time is 17s only.